A Data Dependence Graph in Polaris

Data dependence analysis has become one of the most important constituents of loop-level parallelizing compilers. The information gathered from the analysis is needed to determine both the potential concurrency of loop nests and the legality of loop transformations concerning loop parallelization. The analysis is based on data dependence tests such as the constant test and the GCD test. Once the analysis is complete, the dependence information for the input program is stored for later demand. Since this dependence information is stored in memory, recalculation of the information in following passes is unnecessary, reducing the whole parallelizing compiler execution time. However because this information often requires so much memory space, sometimes up to thousands of Mbytes without strict memory management, it can run out of heap storage, and cause another problem: how do we store the information e ciently? Besides memory space, the time to fetch the necessary information is also another criterion of concern because faster access generally requires more memory space. In this paper, we design an e cient method of dealing with data dependence information in Polaris in terms of memory space and access time. In this method, we use a compromise between the two criteria to get better performance.